[PATCH] shrink-wrap: Once more PRs 67778, 68634, and now 68909

[PATCH] shrink-wrap: Once more PRs 67778, 68634, and now 68909

[Segher Boessenkool]

It turns out v4 wasn't quite complete anyway; so here "v5".

If a candidate PRE cannot get the prologue because a block BB is
reachable from it, but PRE does not dominate BB, we try again with the
dominators of PRE.  That "try again" needs to again consider BB though,
we aren't done with it.

This fixes this problem.  Tested on the 68909 testcase, and bootstrapped
and regression checked on powerpc64-linux.  Is this okay for trunk?


Segher


2015-12-17  Segher Boessenkool  <segher@kernel.crashing.org>

	PR rtl-optimization/67778
	PR rtl-optimization/68634
	PR rtl-optimization/68909
	* shrink-wrap.c (try_shrink_wrapping): If BB isn't dominated by PRE,
	push it back on VEC.

---
 gcc/shrink-wrap.c | 1 +
 1 file changed, 1 insertion(+)

diff --git a/gcc/shrink-wrap.c b/gcc/shrink-wrap.c
index f65b0c3..85e5a8b 100644
--- a/gcc/shrink-wrap.c
+++ b/gcc/shrink-wrap.c
@@ -781,6 +781,7 @@ try_shrink_wrapping (edge *entry_edge, bitmap_head *bb_with,
 	      if (!dominated_by_p (CDI_DOMINATORS, bb, pre))
 		{
 		  ok = false;
+		  vec.quick_push (bb);
 		  break;
 		}
 
-- 
1.9.3

Re: [PATCH] shrink-wrap: Once more PRs 67778, 68634, and now 68909

[Bernd Schmidt]

On 12/17/2015 10:07 PM, Segher Boessenkool wrote:
> It turns out v4 wasn't quite complete anyway; so here "v5".
>
> If a candidate PRE cannot get the prologue because a block BB is
> reachable from it, but PRE does not dominate BB, we try again with the
> dominators of PRE.  That "try again" needs to again consider BB though,
> we aren't done with it.
>
> This fixes this problem.  Tested on the 68909 testcase, and bootstrapped
> and regression checked on powerpc64-linux.  Is this okay for trunk?

This code is getting really quite confusing, and at the least I think we 
need more documentation of what exactly vec is supposed to contain at 
the entry to the inner while loop here.
I'm also beginning to think we should disable this part of the code for 
gcc-6.


Bernd

Re: [PATCH] shrink-wrap: Once more PRs 67778, 68634, and now 68909

[Segher Boessenkool]

On Fri, Dec 18, 2015 at 02:19:37AM +0100, Bernd Schmidt wrote:
> On 12/17/2015 10:07 PM, Segher Boessenkool wrote:
> >It turns out v4 wasn't quite complete anyway; so here "v5".
> >
> >If a candidate PRE cannot get the prologue because a block BB is
> >reachable from it, but PRE does not dominate BB, we try again with the
> >dominators of PRE.  That "try again" needs to again consider BB though,
> >we aren't done with it.
> >
> >This fixes this problem.  Tested on the 68909 testcase, and bootstrapped
> >and regression checked on powerpc64-linux.  Is this okay for trunk?
> 
> This code is getting really quite confusing,

Yes :-(  I don't think stage 3 is the time to completely rewrite it though.

> and at the least I think we 
> need more documentation of what exactly vec is supposed to contain at 
> the entry to the inner while loop here.

Same as in the other loop: vec is a stack of blocks that still need to
be looked at.  I can duplicate the comment if you want?

> I'm also beginning to think we should disable this part of the code for 
> gcc-6.

That would be a regression (from GCC 5); but I understand your worry.
How about we disable it if any further problems show up?


Segher

Re: [PATCH] shrink-wrap: Once more PRs 67778, 68634, and now 68909

[Bernd Schmidt]

On 12/20/2015 05:27 PM, Segher Boessenkool wrote:
> On Fri, Dec 18, 2015 at 02:19:37AM +0100, Bernd Schmidt wrote:
>> On 12/17/2015 10:07 PM, Segher Boessenkool wrote:
>>> It turns out v4 wasn't quite complete anyway; so here "v5".
>>>
>>> If a candidate PRE cannot get the prologue because a block BB is
>>> reachable from it, but PRE does not dominate BB, we try again with the
>>> dominators of PRE.  That "try again" needs to again consider BB though,
>>> we aren't done with it.
>>>
>>> This fixes this problem.  Tested on the 68909 testcase, and bootstrapped
>>> and regression checked on powerpc64-linux.  Is this okay for trunk?
>>
>> This code is getting really quite confusing,
>> and at the least I think we
>> need more documentation of what exactly vec is supposed to contain at
>> the entry to the inner while loop here.
>
> Same as in the other loop: vec is a stack of blocks that still need to
> be looked at.  I can duplicate the comment if you want?

No, I think more is needed. The inner loop looks like it should be 
emptying the vec, but this is not true if we break out of it, and your 
patch now even adds an explicit push. It also looks like it wants to use 
the bb_tmp bitmap to cache results for future iterations of the outer 
loop, but I'm not convinced this is actually correct. I can't follow 
this behaviour anymore without clear a description of intent.

Also, it might be clearer to not modify "pro" in this loop - use a 
"cand" variable, and modify "pro" instead of last_ok, getting rid of the 
latter.

> That would be a regression (from GCC 5); but I understand your worry.
> How about we disable it if any further problems show up?

Let's see whether we can make sense of this code and decide then.


bernd

Re: [PATCH] shrink-wrap: Once more PRs 67778, 68634, and now 68909

[Segher Boessenkool]

On Mon, Jan 04, 2016 at 03:24:10PM +0100, Bernd Schmidt wrote:
> >>This code is getting really quite confusing,
> >>and at the least I think we
> >>need more documentation of what exactly vec is supposed to contain at
> >>the entry to the inner while loop here.
> >
> >Same as in the other loop: vec is a stack of blocks that still need to
> >be looked at.  I can duplicate the comment if you want?
> 
> No, I think more is needed.

Thanks for the review.  New patch attached.

> The inner loop looks like it should be 
> emptying the vec, but this is not true if we break out of it, and your 
> patch now even adds an explicit push.

I added a big comment explaining the algorithm, and changed the push back
to do a peek instead.

> It also looks like it wants to use 
> the bb_tmp bitmap to cache results for future iterations of the outer 
> loop, but I'm not convinced this is actually correct. I can't follow 
> this behaviour anymore without clear a description of intent.

See the new comment.

> Also, it might be clearer to not modify "pro" in this loop - use a 
> "cand" variable, and modify "pro" instead of last_ok, getting rid of the 
> latter.

I tried that, and it doesn't make things clearer in my opinion.  Also
tried assigning "pro = pre" earlier, to make it more similar to the
previous loop; also not an improvement.

The patch also removes a superfluous (and confusing) bitmap set as well
as a bitmap test.

How's this?


Segher


Subject: [PATCH] shrink-wrap: Once more PRs 67778, 68634, and now 68909

If a candidate PRE cannot get the prologue because a block BB is
reachable from it, but PRE does not dominate BB, we try again with the
dominators of PRE.  That "try again" needs to again consider BB though,
we aren't done with it.

This fixes this problem.  Tested on the 68909 testcase, and bootstrapped
and regression checked on powerpc64-linux.  Is this okay for trunk?


2016-01-06  Segher Boessenkool  <segher@kernel.crashing.org>

	PR rtl-optimization/67778
	PR rtl-optimization/68634
	PR rtl-optimization/68909
	* shrink-wrap.c (try_shrink_wrapping): Add comment.  Don't pop
	block from the stack until done with it.  Remove a superfluous
	bitmap set.  Remove a superfluous bitmap test.
---
 gcc/shrink-wrap.c | 23 ++++++++++++++++-------
 1 file changed, 16 insertions(+), 7 deletions(-)

diff --git a/gcc/shrink-wrap.c b/gcc/shrink-wrap.c
index e51bd36..84abd6b 100644
--- a/gcc/shrink-wrap.c
+++ b/gcc/shrink-wrap.c
@@ -750,9 +750,21 @@ try_shrink_wrapping (edge *entry_edge, bitmap_head *bb_with,
 
   /* If we can move PRO back without having to duplicate more blocks, do so.
      We do this because putting the prologue earlier is better for scheduling.
+
      We can move back to a block PRE if every path from PRE will eventually
      need a prologue, that is, PRO is a post-dominator of PRE.  PRE needs
-     to dominate every block reachable from itself.  */
+     to dominate every block reachable from itself.  We keep in BB_TMP a
+     bitmap of the blocks reachable from PRE that we already found, and in
+     VEC a stack of those we still need to consider.
+
+     Any block reachable from PRE is also reachable from all predecessors
+     of PRE, so if we find we need to move PRE back further we can leave
+     everything not considered so far on the stack.  Any block dominated
+     by PRE is also dominated by all other dominators of PRE, so anything
+     found good for some PRE does not need to be reconsidered later.
+
+     We don't need to update BB_WITH because none of the new blocks found
+     can jump to a block that does not need the prologue.  */
 
   if (pro != entry)
     {
@@ -775,18 +787,15 @@ try_shrink_wrapping (edge *entry_edge, bitmap_head *bb_with,
 	  bool ok = true;
 	  while (!vec.is_empty ())
 	    {
-	      basic_block bb = vec.pop ();
-	      bitmap_set_bit (bb_tmp, pre->index);
-
-	      if (!dominated_by_p (CDI_DOMINATORS, bb, pre))
+	      if (!dominated_by_p (CDI_DOMINATORS, vec.last (), pre))
 		{
 		  ok = false;
 		  break;
 		}
 
+	      basic_block bb = vec.pop ();
 	      FOR_EACH_EDGE (e, ei, bb->succs)
-		if (!bitmap_bit_p (bb_with, e->dest->index)
-		    && bitmap_set_bit (bb_tmp, e->dest->index))
+		if (bitmap_set_bit (bb_tmp, e->dest->index))
 		  vec.quick_push (e->dest);
 	    }
 
-- 
1.9.3

Re: [PATCH] shrink-wrap: Once more PRs 67778, 68634, and now 68909

[Segher Boessenkool]

On Wed, Jan 06, 2016 at 06:36:19AM -0600, Segher Boessenkool wrote:
> This fixes this problem.  Tested on the 68909 testcase, and bootstrapped
> and regression checked on powerpc64-linux.  Is this okay for trunk?

Also tested on x86_64-linux now.


Segher


> 2016-01-06  Segher Boessenkool  <segher@kernel.crashing.org>
> 
> 	PR rtl-optimization/67778
> 	PR rtl-optimization/68634
> 	PR rtl-optimization/68909
> 	* shrink-wrap.c (try_shrink_wrapping): Add comment.  Don't pop
> 	block from the stack until done with it.  Remove a superfluous
> 	bitmap set.  Remove a superfluous bitmap test.

Re: [PATCH] shrink-wrap: Once more PRs 67778, 68634, and now 68909

[Bernd Schmidt]

On 01/06/2016 01:36 PM, Segher Boessenkool wrote:
> How's this?
>
> 	PR rtl-optimization/67778
> 	PR rtl-optimization/68634
> 	PR rtl-optimization/68909
> 	* shrink-wrap.c (try_shrink_wrapping): Add comment.  Don't pop
> 	block from the stack until done with it.  Remove a superfluous
> 	bitmap set.  Remove a superfluous bitmap test.

Ok.


Bernd